1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements folding of constants for LLVM. This implements the 10 // (internal) ConstantFold.h interface, which is used by the 11 // ConstantExpr::get* methods to automatically fold constants when possible. 12 // 13 // The current constant folding implementation is implemented in two pieces: the 14 // pieces that don't need DataLayout, and the pieces that do. This is to avoid 15 // a dependence in IR on Target. 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "llvm/IR/ConstantFold.h" 20 #include "llvm/ADT/APSInt.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/IR/Constants.h" 23 #include "llvm/IR/DerivedTypes.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/GetElementPtrTypeIterator.h" 26 #include "llvm/IR/GlobalAlias.h" 27 #include "llvm/IR/GlobalVariable.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/Module.h" 30 #include "llvm/IR/Operator.h" 31 #include "llvm/IR/PatternMatch.h" 32 #include "llvm/Support/ErrorHandling.h" 33 using namespace llvm; 34 using namespace llvm::PatternMatch; 35 36 //===----------------------------------------------------------------------===// 37 // ConstantFold*Instruction Implementations 38 //===----------------------------------------------------------------------===// 39 40 /// Convert the specified vector Constant node to the specified vector type. 41 /// At this point, we know that the elements of the input vector constant are 42 /// all simple integer or FP values. 43 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) { 44 45 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy); 46 if (CV->isNullValue()) return Constant::getNullValue(DstTy); 47 48 // Do not iterate on scalable vector. The num of elements is unknown at 49 // compile-time. 50 if (isa<ScalableVectorType>(DstTy)) 51 return nullptr; 52 53 // If this cast changes element count then we can't handle it here: 54 // doing so requires endianness information. This should be handled by 55 // Analysis/ConstantFolding.cpp 56 unsigned NumElts = cast<FixedVectorType>(DstTy)->getNumElements(); 57 if (NumElts != cast<FixedVectorType>(CV->getType())->getNumElements()) 58 return nullptr; 59 60 Type *DstEltTy = DstTy->getElementType(); 61 // Fast path for splatted constants. 62 if (Constant *Splat = CV->getSplatValue()) { 63 return ConstantVector::getSplat(DstTy->getElementCount(), 64 ConstantExpr::getBitCast(Splat, DstEltTy)); 65 } 66 67 SmallVector<Constant*, 16> Result; 68 Type *Ty = IntegerType::get(CV->getContext(), 32); 69 for (unsigned i = 0; i != NumElts; ++i) { 70 Constant *C = 71 ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i)); 72 C = ConstantExpr::getBitCast(C, DstEltTy); 73 Result.push_back(C); 74 } 75 76 return ConstantVector::get(Result); 77 } 78 79 /// This function determines which opcode to use to fold two constant cast 80 /// expressions together. It uses CastInst::isEliminableCastPair to determine 81 /// the opcode. Consequently its just a wrapper around that function. 82 /// Determine if it is valid to fold a cast of a cast 83 static unsigned 84 foldConstantCastPair( 85 unsigned opc, ///< opcode of the second cast constant expression 86 ConstantExpr *Op, ///< the first cast constant expression 87 Type *DstTy ///< destination type of the first cast 88 ) { 89 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!"); 90 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type"); 91 assert(CastInst::isCast(opc) && "Invalid cast opcode"); 92 93 // The types and opcodes for the two Cast constant expressions 94 Type *SrcTy = Op->getOperand(0)->getType(); 95 Type *MidTy = Op->getType(); 96 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); 97 Instruction::CastOps secondOp = Instruction::CastOps(opc); 98 99 // Assume that pointers are never more than 64 bits wide, and only use this 100 // for the middle type. Otherwise we could end up folding away illegal 101 // bitcasts between address spaces with different sizes. 102 IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext()); 103 104 // Let CastInst::isEliminableCastPair do the heavy lifting. 105 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy, 106 nullptr, FakeIntPtrTy, nullptr); 107 } 108 109 static Constant *FoldBitCast(Constant *V, Type *DestTy) { 110 Type *SrcTy = V->getType(); 111 if (SrcTy == DestTy) 112 return V; // no-op cast 113 114 // Handle casts from one vector constant to another. We know that the src 115 // and dest type have the same size (otherwise its an illegal cast). 116 if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) { 117 if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) { 118 assert(DestPTy->getPrimitiveSizeInBits() == 119 SrcTy->getPrimitiveSizeInBits() && 120 "Not cast between same sized vectors!"); 121 SrcTy = nullptr; 122 // First, check for null. Undef is already handled. 123 if (isa<ConstantAggregateZero>(V)) 124 return Constant::getNullValue(DestTy); 125 126 // Handle ConstantVector and ConstantAggregateVector. 127 return BitCastConstantVector(V, DestPTy); 128 } 129 130 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts 131 // This allows for other simplifications (although some of them 132 // can only be handled by Analysis/ConstantFolding.cpp). 133 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) 134 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy); 135 } 136 137 // Finally, implement bitcast folding now. The code below doesn't handle 138 // bitcast right. 139 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast. 140 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 141 142 // Handle integral constant input. 143 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 144 if (DestTy->isIntegerTy()) 145 // Integral -> Integral. This is a no-op because the bit widths must 146 // be the same. Consequently, we just fold to V. 147 return V; 148 149 // See note below regarding the PPC_FP128 restriction. 150 if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty()) 151 return ConstantFP::get(DestTy->getContext(), 152 APFloat(DestTy->getFltSemantics(), 153 CI->getValue())); 154 155 // Otherwise, can't fold this (vector?) 156 return nullptr; 157 } 158 159 // Handle ConstantFP input: FP -> Integral. 160 if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) { 161 // PPC_FP128 is really the sum of two consecutive doubles, where the first 162 // double is always stored first in memory, regardless of the target 163 // endianness. The memory layout of i128, however, depends on the target 164 // endianness, and so we can't fold this without target endianness 165 // information. This should instead be handled by 166 // Analysis/ConstantFolding.cpp 167 if (FP->getType()->isPPC_FP128Ty()) 168 return nullptr; 169 170 // Make sure dest type is compatible with the folded integer constant. 171 if (!DestTy->isIntegerTy()) 172 return nullptr; 173 174 return ConstantInt::get(FP->getContext(), 175 FP->getValueAPF().bitcastToAPInt()); 176 } 177 178 return nullptr; 179 } 180 181 182 /// V is an integer constant which only has a subset of its bytes used. 183 /// The bytes used are indicated by ByteStart (which is the first byte used, 184 /// counting from the least significant byte) and ByteSize, which is the number 185 /// of bytes used. 186 /// 187 /// This function analyzes the specified constant to see if the specified byte 188 /// range can be returned as a simplified constant. If so, the constant is 189 /// returned, otherwise null is returned. 190 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart, 191 unsigned ByteSize) { 192 assert(C->getType()->isIntegerTy() && 193 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 && 194 "Non-byte sized integer input"); 195 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8; 196 assert(ByteSize && "Must be accessing some piece"); 197 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input"); 198 assert(ByteSize != CSize && "Should not extract everything"); 199 200 // Constant Integers are simple. 201 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 202 APInt V = CI->getValue(); 203 if (ByteStart) 204 V.lshrInPlace(ByteStart*8); 205 V = V.trunc(ByteSize*8); 206 return ConstantInt::get(CI->getContext(), V); 207 } 208 209 // In the input is a constant expr, we might be able to recursively simplify. 210 // If not, we definitely can't do anything. 211 ConstantExpr *CE = dyn_cast<ConstantExpr>(C); 212 if (!CE) return nullptr; 213 214 switch (CE->getOpcode()) { 215 default: return nullptr; 216 case Instruction::Or: { 217 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 218 if (!RHS) 219 return nullptr; 220 221 // X | -1 -> -1. 222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) 223 if (RHSC->isMinusOne()) 224 return RHSC; 225 226 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 227 if (!LHS) 228 return nullptr; 229 return ConstantExpr::getOr(LHS, RHS); 230 } 231 case Instruction::And: { 232 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize); 233 if (!RHS) 234 return nullptr; 235 236 // X & 0 -> 0. 237 if (RHS->isNullValue()) 238 return RHS; 239 240 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize); 241 if (!LHS) 242 return nullptr; 243 return ConstantExpr::getAnd(LHS, RHS); 244 } 245 case Instruction::LShr: { 246 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 247 if (!Amt) 248 return nullptr; 249 APInt ShAmt = Amt->getValue(); 250 // Cannot analyze non-byte shifts. 251 if ((ShAmt & 7) != 0) 252 return nullptr; 253 ShAmt.lshrInPlace(3); 254 255 // If the extract is known to be all zeros, return zero. 256 if (ShAmt.uge(CSize - ByteStart)) 257 return Constant::getNullValue( 258 IntegerType::get(CE->getContext(), ByteSize * 8)); 259 // If the extract is known to be fully in the input, extract it. 260 if (ShAmt.ule(CSize - (ByteStart + ByteSize))) 261 return ExtractConstantBytes(CE->getOperand(0), 262 ByteStart + ShAmt.getZExtValue(), ByteSize); 263 264 // TODO: Handle the 'partially zero' case. 265 return nullptr; 266 } 267 268 case Instruction::Shl: { 269 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1)); 270 if (!Amt) 271 return nullptr; 272 APInt ShAmt = Amt->getValue(); 273 // Cannot analyze non-byte shifts. 274 if ((ShAmt & 7) != 0) 275 return nullptr; 276 ShAmt.lshrInPlace(3); 277 278 // If the extract is known to be all zeros, return zero. 279 if (ShAmt.uge(ByteStart + ByteSize)) 280 return Constant::getNullValue( 281 IntegerType::get(CE->getContext(), ByteSize * 8)); 282 // If the extract is known to be fully in the input, extract it. 283 if (ShAmt.ule(ByteStart)) 284 return ExtractConstantBytes(CE->getOperand(0), 285 ByteStart - ShAmt.getZExtValue(), ByteSize); 286 287 // TODO: Handle the 'partially zero' case. 288 return nullptr; 289 } 290 291 case Instruction::ZExt: { 292 unsigned SrcBitSize = 293 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth(); 294 295 // If extracting something that is completely zero, return 0. 296 if (ByteStart*8 >= SrcBitSize) 297 return Constant::getNullValue(IntegerType::get(CE->getContext(), 298 ByteSize*8)); 299 300 // If exactly extracting the input, return it. 301 if (ByteStart == 0 && ByteSize*8 == SrcBitSize) 302 return CE->getOperand(0); 303 304 // If extracting something completely in the input, if the input is a 305 // multiple of 8 bits, recurse. 306 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize) 307 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize); 308 309 // Otherwise, if extracting a subset of the input, which is not multiple of 310 // 8 bits, do a shift and trunc to get the bits. 311 if ((ByteStart+ByteSize)*8 < SrcBitSize) { 312 assert((SrcBitSize&7) && "Shouldn't get byte sized case here"); 313 Constant *Res = CE->getOperand(0); 314 if (ByteStart) 315 Res = ConstantExpr::getLShr(Res, 316 ConstantInt::get(Res->getType(), ByteStart*8)); 317 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(), 318 ByteSize*8)); 319 } 320 321 // TODO: Handle the 'partially zero' case. 322 return nullptr; 323 } 324 } 325 } 326 327 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, 328 Type *DestTy) { 329 if (isa<PoisonValue>(V)) 330 return PoisonValue::get(DestTy); 331 332 if (isa<UndefValue>(V)) { 333 // zext(undef) = 0, because the top bits will be zero. 334 // sext(undef) = 0, because the top bits will all be the same. 335 // [us]itofp(undef) = 0, because the result value is bounded. 336 if (opc == Instruction::ZExt || opc == Instruction::SExt || 337 opc == Instruction::UIToFP || opc == Instruction::SIToFP) 338 return Constant::getNullValue(DestTy); 339 return UndefValue::get(DestTy); 340 } 341 342 if (V->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() && 343 opc != Instruction::AddrSpaceCast) 344 return Constant::getNullValue(DestTy); 345 346 // If the cast operand is a constant expression, there's a few things we can 347 // do to try to simplify it. 348 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 349 if (CE->isCast()) { 350 // Try hard to fold cast of cast because they are often eliminable. 351 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy)) 352 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy); 353 } else if (CE->getOpcode() == Instruction::GetElementPtr && 354 // Do not fold addrspacecast (gep 0, .., 0). It might make the 355 // addrspacecast uncanonicalized. 356 opc != Instruction::AddrSpaceCast && 357 // Do not fold bitcast (gep) with inrange index, as this loses 358 // information. 359 !cast<GEPOperator>(CE)->getInRangeIndex() && 360 // Do not fold if the gep type is a vector, as bitcasting 361 // operand 0 of a vector gep will result in a bitcast between 362 // different sizes. 363 !CE->getType()->isVectorTy()) { 364 // If all of the indexes in the GEP are null values, there is no pointer 365 // adjustment going on. We might as well cast the source pointer. 366 bool isAllNull = true; 367 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 368 if (!CE->getOperand(i)->isNullValue()) { 369 isAllNull = false; 370 break; 371 } 372 if (isAllNull) 373 // This is casting one pointer type to another, always BitCast 374 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy); 375 } 376 } 377 378 // If the cast operand is a constant vector, perform the cast by 379 // operating on each element. In the cast of bitcasts, the element 380 // count may be mismatched; don't attempt to handle that here. 381 if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) && 382 DestTy->isVectorTy() && 383 cast<FixedVectorType>(DestTy)->getNumElements() == 384 cast<FixedVectorType>(V->getType())->getNumElements()) { 385 VectorType *DestVecTy = cast<VectorType>(DestTy); 386 Type *DstEltTy = DestVecTy->getElementType(); 387 // Fast path for splatted constants. 388 if (Constant *Splat = V->getSplatValue()) { 389 return ConstantVector::getSplat( 390 cast<VectorType>(DestTy)->getElementCount(), 391 ConstantExpr::getCast(opc, Splat, DstEltTy)); 392 } 393 SmallVector<Constant *, 16> res; 394 Type *Ty = IntegerType::get(V->getContext(), 32); 395 for (unsigned i = 0, 396 e = cast<FixedVectorType>(V->getType())->getNumElements(); 397 i != e; ++i) { 398 Constant *C = 399 ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i)); 400 res.push_back(ConstantExpr::getCast(opc, C, DstEltTy)); 401 } 402 return ConstantVector::get(res); 403 } 404 405 // We actually have to do a cast now. Perform the cast according to the 406 // opcode specified. 407 switch (opc) { 408 default: 409 llvm_unreachable("Failed to cast constant expression"); 410 case Instruction::FPTrunc: 411 case Instruction::FPExt: 412 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 413 bool ignored; 414 APFloat Val = FPC->getValueAPF(); 415 Val.convert(DestTy->getFltSemantics(), APFloat::rmNearestTiesToEven, 416 &ignored); 417 return ConstantFP::get(V->getContext(), Val); 418 } 419 return nullptr; // Can't fold. 420 case Instruction::FPToUI: 421 case Instruction::FPToSI: 422 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) { 423 const APFloat &V = FPC->getValueAPF(); 424 bool ignored; 425 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 426 APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI); 427 if (APFloat::opInvalidOp == 428 V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) { 429 // Undefined behavior invoked - the destination type can't represent 430 // the input constant. 431 return PoisonValue::get(DestTy); 432 } 433 return ConstantInt::get(FPC->getContext(), IntVal); 434 } 435 return nullptr; // Can't fold. 436 case Instruction::IntToPtr: //always treated as unsigned 437 if (V->isNullValue()) // Is it an integral null value? 438 return ConstantPointerNull::get(cast<PointerType>(DestTy)); 439 return nullptr; // Other pointer types cannot be casted 440 case Instruction::PtrToInt: // always treated as unsigned 441 // Is it a null pointer value? 442 if (V->isNullValue()) 443 return ConstantInt::get(DestTy, 0); 444 // Other pointer types cannot be casted 445 return nullptr; 446 case Instruction::UIToFP: 447 case Instruction::SIToFP: 448 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 449 const APInt &api = CI->getValue(); 450 APFloat apf(DestTy->getFltSemantics(), 451 APInt::getZero(DestTy->getPrimitiveSizeInBits())); 452 apf.convertFromAPInt(api, opc==Instruction::SIToFP, 453 APFloat::rmNearestTiesToEven); 454 return ConstantFP::get(V->getContext(), apf); 455 } 456 return nullptr; 457 case Instruction::ZExt: 458 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 459 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 460 return ConstantInt::get(V->getContext(), 461 CI->getValue().zext(BitWidth)); 462 } 463 return nullptr; 464 case Instruction::SExt: 465 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 466 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 467 return ConstantInt::get(V->getContext(), 468 CI->getValue().sext(BitWidth)); 469 } 470 return nullptr; 471 case Instruction::Trunc: { 472 if (V->getType()->isVectorTy()) 473 return nullptr; 474 475 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth(); 476 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 477 return ConstantInt::get(V->getContext(), 478 CI->getValue().trunc(DestBitWidth)); 479 } 480 481 // The input must be a constantexpr. See if we can simplify this based on 482 // the bytes we are demanding. Only do this if the source and dest are an 483 // even multiple of a byte. 484 if ((DestBitWidth & 7) == 0 && 485 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0) 486 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8)) 487 return Res; 488 489 return nullptr; 490 } 491 case Instruction::BitCast: 492 return FoldBitCast(V, DestTy); 493 case Instruction::AddrSpaceCast: 494 return nullptr; 495 } 496 } 497 498 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, 499 Constant *V1, Constant *V2) { 500 // Check for i1 and vector true/false conditions. 501 if (Cond->isNullValue()) return V2; 502 if (Cond->isAllOnesValue()) return V1; 503 504 // If the condition is a vector constant, fold the result elementwise. 505 if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) { 506 auto *V1VTy = CondV->getType(); 507 SmallVector<Constant*, 16> Result; 508 Type *Ty = IntegerType::get(CondV->getContext(), 32); 509 for (unsigned i = 0, e = V1VTy->getNumElements(); i != e; ++i) { 510 Constant *V; 511 Constant *V1Element = ConstantExpr::getExtractElement(V1, 512 ConstantInt::get(Ty, i)); 513 Constant *V2Element = ConstantExpr::getExtractElement(V2, 514 ConstantInt::get(Ty, i)); 515 auto *Cond = cast<Constant>(CondV->getOperand(i)); 516 if (isa<PoisonValue>(Cond)) { 517 V = PoisonValue::get(V1Element->getType()); 518 } else if (V1Element == V2Element) { 519 V = V1Element; 520 } else if (isa<UndefValue>(Cond)) { 521 V = isa<UndefValue>(V1Element) ? V1Element : V2Element; 522 } else { 523 if (!isa<ConstantInt>(Cond)) break; 524 V = Cond->isNullValue() ? V2Element : V1Element; 525 } 526 Result.push_back(V); 527 } 528 529 // If we were able to build the vector, return it. 530 if (Result.size() == V1VTy->getNumElements()) 531 return ConstantVector::get(Result); 532 } 533 534 if (isa<PoisonValue>(Cond)) 535 return PoisonValue::get(V1->getType()); 536 537 if (isa<UndefValue>(Cond)) { 538 if (isa<UndefValue>(V1)) return V1; 539 return V2; 540 } 541 542 if (V1 == V2) return V1; 543 544 if (isa<PoisonValue>(V1)) 545 return V2; 546 if (isa<PoisonValue>(V2)) 547 return V1; 548 549 // If the true or false value is undef, we can fold to the other value as 550 // long as the other value isn't poison. 551 auto NotPoison = [](Constant *C) { 552 if (isa<PoisonValue>(C)) 553 return false; 554 555 // TODO: We can analyze ConstExpr by opcode to determine if there is any 556 // possibility of poison. 557 if (isa<ConstantExpr>(C)) 558 return false; 559 560 if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(C) || 561 isa<ConstantPointerNull>(C) || isa<Function>(C)) 562 return true; 563 564 if (C->getType()->isVectorTy()) 565 return !C->containsPoisonElement() && !C->containsConstantExpression(); 566 567 // TODO: Recursively analyze aggregates or other constants. 568 return false; 569 }; 570 if (isa<UndefValue>(V1) && NotPoison(V2)) return V2; 571 if (isa<UndefValue>(V2) && NotPoison(V1)) return V1; 572 573 return nullptr; 574 } 575 576 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val, 577 Constant *Idx) { 578 auto *ValVTy = cast<VectorType>(Val->getType()); 579 580 // extractelt poison, C -> poison 581 // extractelt C, undef -> poison 582 if (isa<PoisonValue>(Val) || isa<UndefValue>(Idx)) 583 return PoisonValue::get(ValVTy->getElementType()); 584 585 // extractelt undef, C -> undef 586 if (isa<UndefValue>(Val)) 587 return UndefValue::get(ValVTy->getElementType()); 588 589 auto *CIdx = dyn_cast<ConstantInt>(Idx); 590 if (!CIdx) 591 return nullptr; 592 593 if (auto *ValFVTy = dyn_cast<FixedVectorType>(Val->getType())) { 594 // ee({w,x,y,z}, wrong_value) -> poison 595 if (CIdx->uge(ValFVTy->getNumElements())) 596 return PoisonValue::get(ValFVTy->getElementType()); 597 } 598 599 // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...) 600 if (auto *CE = dyn_cast<ConstantExpr>(Val)) { 601 if (auto *GEP = dyn_cast<GEPOperator>(CE)) { 602 SmallVector<Constant *, 8> Ops; 603 Ops.reserve(CE->getNumOperands()); 604 for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) { 605 Constant *Op = CE->getOperand(i); 606 if (Op->getType()->isVectorTy()) { 607 Constant *ScalarOp = ConstantExpr::getExtractElement(Op, Idx); 608 if (!ScalarOp) 609 return nullptr; 610 Ops.push_back(ScalarOp); 611 } else 612 Ops.push_back(Op); 613 } 614 return CE->getWithOperands(Ops, ValVTy->getElementType(), false, 615 GEP->getSourceElementType()); 616 } else if (CE->getOpcode() == Instruction::InsertElement) { 617 if (const auto *IEIdx = dyn_cast<ConstantInt>(CE->getOperand(2))) { 618 if (APSInt::isSameValue(APSInt(IEIdx->getValue()), 619 APSInt(CIdx->getValue()))) { 620 return CE->getOperand(1); 621 } else { 622 return ConstantExpr::getExtractElement(CE->getOperand(0), CIdx); 623 } 624 } 625 } 626 } 627 628 if (Constant *C = Val->getAggregateElement(CIdx)) 629 return C; 630 631 // Lane < Splat minimum vector width => extractelt Splat(x), Lane -> x 632 if (CIdx->getValue().ult(ValVTy->getElementCount().getKnownMinValue())) { 633 if (Constant *SplatVal = Val->getSplatValue()) 634 return SplatVal; 635 } 636 637 return nullptr; 638 } 639 640 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, 641 Constant *Elt, 642 Constant *Idx) { 643 if (isa<UndefValue>(Idx)) 644 return PoisonValue::get(Val->getType()); 645 646 // Inserting null into all zeros is still all zeros. 647 // TODO: This is true for undef and poison splats too. 648 if (isa<ConstantAggregateZero>(Val) && Elt->isNullValue()) 649 return Val; 650 651 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx); 652 if (!CIdx) return nullptr; 653 654 // Do not iterate on scalable vector. The num of elements is unknown at 655 // compile-time. 656 if (isa<ScalableVectorType>(Val->getType())) 657 return nullptr; 658 659 auto *ValTy = cast<FixedVectorType>(Val->getType()); 660 661 unsigned NumElts = ValTy->getNumElements(); 662 if (CIdx->uge(NumElts)) 663 return PoisonValue::get(Val->getType()); 664 665 SmallVector<Constant*, 16> Result; 666 Result.reserve(NumElts); 667 auto *Ty = Type::getInt32Ty(Val->getContext()); 668 uint64_t IdxVal = CIdx->getZExtValue(); 669 for (unsigned i = 0; i != NumElts; ++i) { 670 if (i == IdxVal) { 671 Result.push_back(Elt); 672 continue; 673 } 674 675 Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i)); 676 Result.push_back(C); 677 } 678 679 return ConstantVector::get(Result); 680 } 681 682 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2, 683 ArrayRef<int> Mask) { 684 auto *V1VTy = cast<VectorType>(V1->getType()); 685 unsigned MaskNumElts = Mask.size(); 686 auto MaskEltCount = 687 ElementCount::get(MaskNumElts, isa<ScalableVectorType>(V1VTy)); 688 Type *EltTy = V1VTy->getElementType(); 689 690 // Poison shuffle mask -> poison value. 691 if (all_of(Mask, [](int Elt) { return Elt == PoisonMaskElem; })) { 692 return PoisonValue::get(VectorType::get(EltTy, MaskEltCount)); 693 } 694 695 // If the mask is all zeros this is a splat, no need to go through all 696 // elements. 697 if (all_of(Mask, [](int Elt) { return Elt == 0; })) { 698 Type *Ty = IntegerType::get(V1->getContext(), 32); 699 Constant *Elt = 700 ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, 0)); 701 702 if (Elt->isNullValue()) { 703 auto *VTy = VectorType::get(EltTy, MaskEltCount); 704 return ConstantAggregateZero::get(VTy); 705 } else if (!MaskEltCount.isScalable()) 706 return ConstantVector::getSplat(MaskEltCount, Elt); 707 } 708 // Do not iterate on scalable vector. The num of elements is unknown at 709 // compile-time. 710 if (isa<ScalableVectorType>(V1VTy)) 711 return nullptr; 712 713 unsigned SrcNumElts = V1VTy->getElementCount().getKnownMinValue(); 714 715 // Loop over the shuffle mask, evaluating each element. 716 SmallVector<Constant*, 32> Result; 717 for (unsigned i = 0; i != MaskNumElts; ++i) { 718 int Elt = Mask[i]; 719 if (Elt == -1) { 720 Result.push_back(UndefValue::get(EltTy)); 721 continue; 722 } 723 Constant *InElt; 724 if (unsigned(Elt) >= SrcNumElts*2) 725 InElt = UndefValue::get(EltTy); 726 else if (unsigned(Elt) >= SrcNumElts) { 727 Type *Ty = IntegerType::get(V2->getContext(), 32); 728 InElt = 729 ConstantExpr::getExtractElement(V2, 730 ConstantInt::get(Ty, Elt - SrcNumElts)); 731 } else { 732 Type *Ty = IntegerType::get(V1->getContext(), 32); 733 InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt)); 734 } 735 Result.push_back(InElt); 736 } 737 738 return ConstantVector::get(Result); 739 } 740 741 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg, 742 ArrayRef<unsigned> Idxs) { 743 // Base case: no indices, so return the entire value. 744 if (Idxs.empty()) 745 return Agg; 746 747 if (Constant *C = Agg->getAggregateElement(Idxs[0])) 748 return ConstantFoldExtractValueInstruction(C, Idxs.slice(1)); 749 750 return nullptr; 751 } 752 753 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, 754 Constant *Val, 755 ArrayRef<unsigned> Idxs) { 756 // Base case: no indices, so replace the entire value. 757 if (Idxs.empty()) 758 return Val; 759 760 unsigned NumElts; 761 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 762 NumElts = ST->getNumElements(); 763 else 764 NumElts = cast<ArrayType>(Agg->getType())->getNumElements(); 765 766 SmallVector<Constant*, 32> Result; 767 for (unsigned i = 0; i != NumElts; ++i) { 768 Constant *C = Agg->getAggregateElement(i); 769 if (!C) return nullptr; 770 771 if (Idxs[0] == i) 772 C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1)); 773 774 Result.push_back(C); 775 } 776 777 if (StructType *ST = dyn_cast<StructType>(Agg->getType())) 778 return ConstantStruct::get(ST, Result); 779 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Result); 780 } 781 782 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) { 783 assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected"); 784 785 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length 786 // vectors are always evaluated per element. 787 bool IsScalableVector = isa<ScalableVectorType>(C->getType()); 788 bool HasScalarUndefOrScalableVectorUndef = 789 (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(C); 790 791 if (HasScalarUndefOrScalableVectorUndef) { 792 switch (static_cast<Instruction::UnaryOps>(Opcode)) { 793 case Instruction::FNeg: 794 return C; // -undef -> undef 795 case Instruction::UnaryOpsEnd: 796 llvm_unreachable("Invalid UnaryOp"); 797 } 798 } 799 800 // Constant should not be UndefValue, unless these are vector constants. 801 assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue"); 802 // We only have FP UnaryOps right now. 803 assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp"); 804 805 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 806 const APFloat &CV = CFP->getValueAPF(); 807 switch (Opcode) { 808 default: 809 break; 810 case Instruction::FNeg: 811 return ConstantFP::get(C->getContext(), neg(CV)); 812 } 813 } else if (auto *VTy = dyn_cast<FixedVectorType>(C->getType())) { 814 815 Type *Ty = IntegerType::get(VTy->getContext(), 32); 816 // Fast path for splatted constants. 817 if (Constant *Splat = C->getSplatValue()) 818 if (Constant *Elt = ConstantFoldUnaryInstruction(Opcode, Splat)) 819 return ConstantVector::getSplat(VTy->getElementCount(), Elt); 820 821 // Fold each element and create a vector constant from those constants. 822 SmallVector<Constant *, 16> Result; 823 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { 824 Constant *ExtractIdx = ConstantInt::get(Ty, i); 825 Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx); 826 Constant *Res = ConstantFoldUnaryInstruction(Opcode, Elt); 827 if (!Res) 828 return nullptr; 829 Result.push_back(Res); 830 } 831 832 return ConstantVector::get(Result); 833 } 834 835 // We don't know how to fold this. 836 return nullptr; 837 } 838 839 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1, 840 Constant *C2) { 841 assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected"); 842 843 // Simplify BinOps with their identity values first. They are no-ops and we 844 // can always return the other value, including undef or poison values. 845 // FIXME: remove unnecessary duplicated identity patterns below. 846 // FIXME: Use AllowRHSConstant with getBinOpIdentity to handle additional ops, 847 // like X << 0 = X. 848 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, C1->getType()); 849 if (Identity) { 850 if (C1 == Identity) 851 return C2; 852 if (C2 == Identity) 853 return C1; 854 } 855 856 // Binary operations propagate poison. 857 if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2)) 858 return PoisonValue::get(C1->getType()); 859 860 // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length 861 // vectors are always evaluated per element. 862 bool IsScalableVector = isa<ScalableVectorType>(C1->getType()); 863 bool HasScalarUndefOrScalableVectorUndef = 864 (!C1->getType()->isVectorTy() || IsScalableVector) && 865 (isa<UndefValue>(C1) || isa<UndefValue>(C2)); 866 if (HasScalarUndefOrScalableVectorUndef) { 867 switch (static_cast<Instruction::BinaryOps>(Opcode)) { 868 case Instruction::Xor: 869 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 870 // Handle undef ^ undef -> 0 special case. This is a common 871 // idiom (misuse). 872 return Constant::getNullValue(C1->getType()); 873 [[fallthrough]]; 874 case Instruction::Add: 875 case Instruction::Sub: 876 return UndefValue::get(C1->getType()); 877 case Instruction::And: 878 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef 879 return C1; 880 return Constant::getNullValue(C1->getType()); // undef & X -> 0 881 case Instruction::Mul: { 882 // undef * undef -> undef 883 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 884 return C1; 885 const APInt *CV; 886 // X * undef -> undef if X is odd 887 if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV))) 888 if ((*CV)[0]) 889 return UndefValue::get(C1->getType()); 890 891 // X * undef -> 0 otherwise 892 return Constant::getNullValue(C1->getType()); 893 } 894 case Instruction::SDiv: 895 case Instruction::UDiv: 896 // X / undef -> poison 897 // X / 0 -> poison 898 if (match(C2, m_CombineOr(m_Undef(), m_Zero()))) 899 return PoisonValue::get(C2->getType()); 900 // undef / 1 -> undef 901 if (match(C2, m_One())) 902 return C1; 903 // undef / X -> 0 otherwise 904 return Constant::getNullValue(C1->getType()); 905 case Instruction::URem: 906 case Instruction::SRem: 907 // X % undef -> poison 908 // X % 0 -> poison 909 if (match(C2, m_CombineOr(m_Undef(), m_Zero()))) 910 return PoisonValue::get(C2->getType()); 911 // undef % X -> 0 otherwise 912 return Constant::getNullValue(C1->getType()); 913 case Instruction::Or: // X | undef -> -1 914 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef 915 return C1; 916 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0 917 case Instruction::LShr: 918 // X >>l undef -> poison 919 if (isa<UndefValue>(C2)) 920 return PoisonValue::get(C2->getType()); 921 // undef >>l 0 -> undef 922 if (match(C2, m_Zero())) 923 return C1; 924 // undef >>l X -> 0 925 return Constant::getNullValue(C1->getType()); 926 case Instruction::AShr: 927 // X >>a undef -> poison 928 if (isa<UndefValue>(C2)) 929 return PoisonValue::get(C2->getType()); 930 // undef >>a 0 -> undef 931 if (match(C2, m_Zero())) 932 return C1; 933 // TODO: undef >>a X -> poison if the shift is exact 934 // undef >>a X -> 0 935 return Constant::getNullValue(C1->getType()); 936 case Instruction::Shl: 937 // X << undef -> undef 938 if (isa<UndefValue>(C2)) 939 return PoisonValue::get(C2->getType()); 940 // undef << 0 -> undef 941 if (match(C2, m_Zero())) 942 return C1; 943 // undef << X -> 0 944 return Constant::getNullValue(C1->getType()); 945 case Instruction::FSub: 946 // -0.0 - undef --> undef (consistent with "fneg undef") 947 if (match(C1, m_NegZeroFP()) && isa<UndefValue>(C2)) 948 return C2; 949 [[fallthrough]]; 950 case Instruction::FAdd: 951 case Instruction::FMul: 952 case Instruction::FDiv: 953 case Instruction::FRem: 954 // [any flop] undef, undef -> undef 955 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) 956 return C1; 957 // [any flop] C, undef -> NaN 958 // [any flop] undef, C -> NaN 959 // We could potentially specialize NaN/Inf constants vs. 'normal' 960 // constants (possibly differently depending on opcode and operand). This 961 // would allow returning undef sometimes. But it is always safe to fold to 962 // NaN because we can choose the undef operand as NaN, and any FP opcode 963 // with a NaN operand will propagate NaN. 964 return ConstantFP::getNaN(C1->getType()); 965 case Instruction::BinaryOpsEnd: 966 llvm_unreachable("Invalid BinaryOp"); 967 } 968 } 969 970 // Neither constant should be UndefValue, unless these are vector constants. 971 assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue"); 972 973 // Handle simplifications when the RHS is a constant int. 974 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 975 switch (Opcode) { 976 case Instruction::Add: 977 if (CI2->isZero()) return C1; // X + 0 == X 978 break; 979 case Instruction::Sub: 980 if (CI2->isZero()) return C1; // X - 0 == X 981 break; 982 case Instruction::Mul: 983 if (CI2->isZero()) return C2; // X * 0 == 0 984 if (CI2->isOne()) 985 return C1; // X * 1 == X 986 break; 987 case Instruction::UDiv: 988 case Instruction::SDiv: 989 if (CI2->isOne()) 990 return C1; // X / 1 == X 991 if (CI2->isZero()) 992 return PoisonValue::get(CI2->getType()); // X / 0 == poison 993 break; 994 case Instruction::URem: 995 case Instruction::SRem: 996 if (CI2->isOne()) 997 return Constant::getNullValue(CI2->getType()); // X % 1 == 0 998 if (CI2->isZero()) 999 return PoisonValue::get(CI2->getType()); // X % 0 == poison 1000 break; 1001 case Instruction::And: 1002 if (CI2->isZero()) return C2; // X & 0 == 0 1003 if (CI2->isMinusOne()) 1004 return C1; // X & -1 == X 1005 1006 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1007 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64) 1008 if (CE1->getOpcode() == Instruction::ZExt) { 1009 unsigned DstWidth = CI2->getType()->getBitWidth(); 1010 unsigned SrcWidth = 1011 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits(); 1012 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1013 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits) 1014 return C1; 1015 } 1016 1017 // If and'ing the address of a global with a constant, fold it. 1018 if (CE1->getOpcode() == Instruction::PtrToInt && 1019 isa<GlobalValue>(CE1->getOperand(0))) { 1020 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0)); 1021 1022 Align GVAlign; // defaults to 1 1023 1024 if (Module *TheModule = GV->getParent()) { 1025 const DataLayout &DL = TheModule->getDataLayout(); 1026 GVAlign = GV->getPointerAlignment(DL); 1027 1028 // If the function alignment is not specified then assume that it 1029 // is 4. 1030 // This is dangerous; on x86, the alignment of the pointer 1031 // corresponds to the alignment of the function, but might be less 1032 // than 4 if it isn't explicitly specified. 1033 // However, a fix for this behaviour was reverted because it 1034 // increased code size (see https://reviews.llvm.org/D55115) 1035 // FIXME: This code should be deleted once existing targets have 1036 // appropriate defaults 1037 if (isa<Function>(GV) && !DL.getFunctionPtrAlign()) 1038 GVAlign = Align(4); 1039 } else if (isa<GlobalVariable>(GV)) { 1040 GVAlign = cast<GlobalVariable>(GV)->getAlign().valueOrOne(); 1041 } 1042 1043 if (GVAlign > 1) { 1044 unsigned DstWidth = CI2->getType()->getBitWidth(); 1045 unsigned SrcWidth = std::min(DstWidth, Log2(GVAlign)); 1046 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth)); 1047 1048 // If checking bits we know are clear, return zero. 1049 if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) 1050 return Constant::getNullValue(CI2->getType()); 1051 } 1052 } 1053 } 1054 break; 1055 case Instruction::Or: 1056 if (CI2->isZero()) return C1; // X | 0 == X 1057 if (CI2->isMinusOne()) 1058 return C2; // X | -1 == -1 1059 break; 1060 case Instruction::Xor: 1061 if (CI2->isZero()) return C1; // X ^ 0 == X 1062 1063 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1064 switch (CE1->getOpcode()) { 1065 default: break; 1066 case Instruction::ICmp: 1067 case Instruction::FCmp: 1068 // cmp pred ^ true -> cmp !pred 1069 assert(CI2->isOne()); 1070 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate(); 1071 pred = CmpInst::getInversePredicate(pred); 1072 return ConstantExpr::getCompare(pred, CE1->getOperand(0), 1073 CE1->getOperand(1)); 1074 } 1075 } 1076 break; 1077 case Instruction::AShr: 1078 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2 1079 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) 1080 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero. 1081 return ConstantExpr::getLShr(C1, C2); 1082 break; 1083 } 1084 } else if (isa<ConstantInt>(C1)) { 1085 // If C1 is a ConstantInt and C2 is not, swap the operands. 1086 if (Instruction::isCommutative(Opcode)) 1087 return ConstantExpr::get(Opcode, C2, C1); 1088 } 1089 1090 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) { 1091 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) { 1092 const APInt &C1V = CI1->getValue(); 1093 const APInt &C2V = CI2->getValue(); 1094 switch (Opcode) { 1095 default: 1096 break; 1097 case Instruction::Add: 1098 return ConstantInt::get(CI1->getContext(), C1V + C2V); 1099 case Instruction::Sub: 1100 return ConstantInt::get(CI1->getContext(), C1V - C2V); 1101 case Instruction::Mul: 1102 return ConstantInt::get(CI1->getContext(), C1V * C2V); 1103 case Instruction::UDiv: 1104 assert(!CI2->isZero() && "Div by zero handled above"); 1105 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V)); 1106 case Instruction::SDiv: 1107 assert(!CI2->isZero() && "Div by zero handled above"); 1108 if (C2V.isAllOnes() && C1V.isMinSignedValue()) 1109 return PoisonValue::get(CI1->getType()); // MIN_INT / -1 -> poison 1110 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V)); 1111 case Instruction::URem: 1112 assert(!CI2->isZero() && "Div by zero handled above"); 1113 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V)); 1114 case Instruction::SRem: 1115 assert(!CI2->isZero() && "Div by zero handled above"); 1116 if (C2V.isAllOnes() && C1V.isMinSignedValue()) 1117 return PoisonValue::get(CI1->getType()); // MIN_INT % -1 -> poison 1118 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V)); 1119 case Instruction::And: 1120 return ConstantInt::get(CI1->getContext(), C1V & C2V); 1121 case Instruction::Or: 1122 return ConstantInt::get(CI1->getContext(), C1V | C2V); 1123 case Instruction::Xor: 1124 return ConstantInt::get(CI1->getContext(), C1V ^ C2V); 1125 case Instruction::Shl: 1126 if (C2V.ult(C1V.getBitWidth())) 1127 return ConstantInt::get(CI1->getContext(), C1V.shl(C2V)); 1128 return PoisonValue::get(C1->getType()); // too big shift is poison 1129 case Instruction::LShr: 1130 if (C2V.ult(C1V.getBitWidth())) 1131 return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V)); 1132 return PoisonValue::get(C1->getType()); // too big shift is poison 1133 case Instruction::AShr: 1134 if (C2V.ult(C1V.getBitWidth())) 1135 return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V)); 1136 return PoisonValue::get(C1->getType()); // too big shift is poison 1137 } 1138 } 1139 1140 switch (Opcode) { 1141 case Instruction::SDiv: 1142 case Instruction::UDiv: 1143 case Instruction::URem: 1144 case Instruction::SRem: 1145 case Instruction::LShr: 1146 case Instruction::AShr: 1147 case Instruction::Shl: 1148 if (CI1->isZero()) return C1; 1149 break; 1150 default: 1151 break; 1152 } 1153 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) { 1154 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) { 1155 const APFloat &C1V = CFP1->getValueAPF(); 1156 const APFloat &C2V = CFP2->getValueAPF(); 1157 APFloat C3V = C1V; // copy for modification 1158 switch (Opcode) { 1159 default: 1160 break; 1161 case Instruction::FAdd: 1162 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven); 1163 return ConstantFP::get(C1->getContext(), C3V); 1164 case Instruction::FSub: 1165 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven); 1166 return ConstantFP::get(C1->getContext(), C3V); 1167 case Instruction::FMul: 1168 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven); 1169 return ConstantFP::get(C1->getContext(), C3V); 1170 case Instruction::FDiv: 1171 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven); 1172 return ConstantFP::get(C1->getContext(), C3V); 1173 case Instruction::FRem: 1174 (void)C3V.mod(C2V); 1175 return ConstantFP::get(C1->getContext(), C3V); 1176 } 1177 } 1178 } else if (auto *VTy = dyn_cast<VectorType>(C1->getType())) { 1179 // Fast path for splatted constants. 1180 if (Constant *C2Splat = C2->getSplatValue()) { 1181 if (Instruction::isIntDivRem(Opcode) && C2Splat->isNullValue()) 1182 return PoisonValue::get(VTy); 1183 if (Constant *C1Splat = C1->getSplatValue()) { 1184 Constant *Res = 1185 ConstantExpr::isDesirableBinOp(Opcode) 1186 ? ConstantExpr::get(Opcode, C1Splat, C2Splat) 1187 : ConstantFoldBinaryInstruction(Opcode, C1Splat, C2Splat); 1188 if (!Res) 1189 return nullptr; 1190 return ConstantVector::getSplat(VTy->getElementCount(), Res); 1191 } 1192 } 1193 1194 if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) { 1195 // Fold each element and create a vector constant from those constants. 1196 SmallVector<Constant*, 16> Result; 1197 Type *Ty = IntegerType::get(FVTy->getContext(), 32); 1198 for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) { 1199 Constant *ExtractIdx = ConstantInt::get(Ty, i); 1200 Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx); 1201 Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx); 1202 1203 // If any element of a divisor vector is zero, the whole op is poison. 1204 if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue()) 1205 return PoisonValue::get(VTy); 1206 1207 Constant *Res = ConstantExpr::isDesirableBinOp(Opcode) 1208 ? ConstantExpr::get(Opcode, LHS, RHS) 1209 : ConstantFoldBinaryInstruction(Opcode, LHS, RHS); 1210 if (!Res) 1211 return nullptr; 1212 Result.push_back(Res); 1213 } 1214 1215 return ConstantVector::get(Result); 1216 } 1217 } 1218 1219 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1220 // There are many possible foldings we could do here. We should probably 1221 // at least fold add of a pointer with an integer into the appropriate 1222 // getelementptr. This will improve alias analysis a bit. 1223 1224 // Given ((a + b) + c), if (b + c) folds to something interesting, return 1225 // (a + (b + c)). 1226 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { 1227 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2); 1228 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode) 1229 return ConstantExpr::get(Opcode, CE1->getOperand(0), T); 1230 } 1231 } else if (isa<ConstantExpr>(C2)) { 1232 // If C2 is a constant expr and C1 isn't, flop them around and fold the 1233 // other way if possible. 1234 if (Instruction::isCommutative(Opcode)) 1235 return ConstantFoldBinaryInstruction(Opcode, C2, C1); 1236 } 1237 1238 // i1 can be simplified in many cases. 1239 if (C1->getType()->isIntegerTy(1)) { 1240 switch (Opcode) { 1241 case Instruction::Add: 1242 case Instruction::Sub: 1243 return ConstantExpr::getXor(C1, C2); 1244 case Instruction::Mul: 1245 return ConstantExpr::getAnd(C1, C2); 1246 case Instruction::Shl: 1247 case Instruction::LShr: 1248 case Instruction::AShr: 1249 // We can assume that C2 == 0. If it were one the result would be 1250 // undefined because the shift value is as large as the bitwidth. 1251 return C1; 1252 case Instruction::SDiv: 1253 case Instruction::UDiv: 1254 // We can assume that C2 == 1. If it were zero the result would be 1255 // undefined through division by zero. 1256 return C1; 1257 case Instruction::URem: 1258 case Instruction::SRem: 1259 // We can assume that C2 == 1. If it were zero the result would be 1260 // undefined through division by zero. 1261 return ConstantInt::getFalse(C1->getContext()); 1262 default: 1263 break; 1264 } 1265 } 1266 1267 // We don't know how to fold this. 1268 return nullptr; 1269 } 1270 1271 /// This function determines if there is anything we can decide about the two 1272 /// constants provided. This doesn't need to handle simple things like 1273 /// ConstantFP comparisons, but should instead handle ConstantExprs. 1274 /// If we can determine that the two constants have a particular relation to 1275 /// each other, we should return the corresponding FCmpInst predicate, 1276 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in 1277 /// ConstantFoldCompareInstruction. 1278 /// 1279 /// To simplify this code we canonicalize the relation so that the first 1280 /// operand is always the most "complex" of the two. We consider ConstantFP 1281 /// to be the simplest, and ConstantExprs to be the most complex. 1282 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) { 1283 assert(V1->getType() == V2->getType() && 1284 "Cannot compare values of different types!"); 1285 1286 // We do not know if a constant expression will evaluate to a number or NaN. 1287 // Therefore, we can only say that the relation is unordered or equal. 1288 if (V1 == V2) return FCmpInst::FCMP_UEQ; 1289 1290 if (!isa<ConstantExpr>(V1)) { 1291 if (!isa<ConstantExpr>(V2)) { 1292 // Simple case, use the standard constant folder. 1293 ConstantInt *R = nullptr; 1294 R = dyn_cast<ConstantInt>( 1295 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2)); 1296 if (R && !R->isZero()) 1297 return FCmpInst::FCMP_OEQ; 1298 R = dyn_cast<ConstantInt>( 1299 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2)); 1300 if (R && !R->isZero()) 1301 return FCmpInst::FCMP_OLT; 1302 R = dyn_cast<ConstantInt>( 1303 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2)); 1304 if (R && !R->isZero()) 1305 return FCmpInst::FCMP_OGT; 1306 1307 // Nothing more we can do 1308 return FCmpInst::BAD_FCMP_PREDICATE; 1309 } 1310 1311 // If the first operand is simple and second is ConstantExpr, swap operands. 1312 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1); 1313 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE) 1314 return FCmpInst::getSwappedPredicate(SwappedRelation); 1315 } else { 1316 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1317 // constantexpr or a simple constant. 1318 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1319 switch (CE1->getOpcode()) { 1320 case Instruction::FPTrunc: 1321 case Instruction::FPExt: 1322 case Instruction::UIToFP: 1323 case Instruction::SIToFP: 1324 // We might be able to do something with these but we don't right now. 1325 break; 1326 default: 1327 break; 1328 } 1329 } 1330 // There are MANY other foldings that we could perform here. They will 1331 // probably be added on demand, as they seem needed. 1332 return FCmpInst::BAD_FCMP_PREDICATE; 1333 } 1334 1335 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, 1336 const GlobalValue *GV2) { 1337 auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { 1338 if (GV->isInterposable() || GV->hasGlobalUnnamedAddr()) 1339 return true; 1340 if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) { 1341 Type *Ty = GVar->getValueType(); 1342 // A global with opaque type might end up being zero sized. 1343 if (!Ty->isSized()) 1344 return true; 1345 // A global with an empty type might lie at the address of any other 1346 // global. 1347 if (Ty->isEmptyTy()) 1348 return true; 1349 } 1350 return false; 1351 }; 1352 // Don't try to decide equality of aliases. 1353 if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2)) 1354 if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) 1355 return ICmpInst::ICMP_NE; 1356 return ICmpInst::BAD_ICMP_PREDICATE; 1357 } 1358 1359 /// This function determines if there is anything we can decide about the two 1360 /// constants provided. This doesn't need to handle simple things like integer 1361 /// comparisons, but should instead handle ConstantExprs and GlobalValues. 1362 /// If we can determine that the two constants have a particular relation to 1363 /// each other, we should return the corresponding ICmp predicate, otherwise 1364 /// return ICmpInst::BAD_ICMP_PREDICATE. 1365 /// 1366 /// To simplify this code we canonicalize the relation so that the first 1367 /// operand is always the most "complex" of the two. We consider simple 1368 /// constants (like ConstantInt) to be the simplest, followed by 1369 /// GlobalValues, followed by ConstantExpr's (the most complex). 1370 /// 1371 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2, 1372 bool isSigned) { 1373 assert(V1->getType() == V2->getType() && 1374 "Cannot compare different types of values!"); 1375 if (V1 == V2) return ICmpInst::ICMP_EQ; 1376 1377 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) && 1378 !isa<BlockAddress>(V1)) { 1379 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) && 1380 !isa<BlockAddress>(V2)) { 1381 // We distilled this down to a simple case, use the standard constant 1382 // folder. 1383 ConstantInt *R = nullptr; 1384 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ; 1385 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1386 if (R && !R->isZero()) 1387 return pred; 1388 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1389 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1390 if (R && !R->isZero()) 1391 return pred; 1392 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1393 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2)); 1394 if (R && !R->isZero()) 1395 return pred; 1396 1397 // If we couldn't figure it out, bail. 1398 return ICmpInst::BAD_ICMP_PREDICATE; 1399 } 1400 1401 // If the first operand is simple, swap operands. 1402 ICmpInst::Predicate SwappedRelation = 1403 evaluateICmpRelation(V2, V1, isSigned); 1404 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1405 return ICmpInst::getSwappedPredicate(SwappedRelation); 1406 1407 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) { 1408 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1409 ICmpInst::Predicate SwappedRelation = 1410 evaluateICmpRelation(V2, V1, isSigned); 1411 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1412 return ICmpInst::getSwappedPredicate(SwappedRelation); 1413 return ICmpInst::BAD_ICMP_PREDICATE; 1414 } 1415 1416 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1417 // constant (which, since the types must match, means that it's a 1418 // ConstantPointerNull). 1419 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1420 return areGlobalsPotentiallyEqual(GV, GV2); 1421 } else if (isa<BlockAddress>(V2)) { 1422 return ICmpInst::ICMP_NE; // Globals never equal labels. 1423 } else { 1424 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!"); 1425 // GlobalVals can never be null unless they have external weak linkage. 1426 // We don't try to evaluate aliases here. 1427 // NOTE: We should not be doing this constant folding if null pointer 1428 // is considered valid for the function. But currently there is no way to 1429 // query it from the Constant type. 1430 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) && 1431 !NullPointerIsDefined(nullptr /* F */, 1432 GV->getType()->getAddressSpace())) 1433 return ICmpInst::ICMP_UGT; 1434 } 1435 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) { 1436 if (isa<ConstantExpr>(V2)) { // Swap as necessary. 1437 ICmpInst::Predicate SwappedRelation = 1438 evaluateICmpRelation(V2, V1, isSigned); 1439 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) 1440 return ICmpInst::getSwappedPredicate(SwappedRelation); 1441 return ICmpInst::BAD_ICMP_PREDICATE; 1442 } 1443 1444 // Now we know that the RHS is a GlobalValue, BlockAddress or simple 1445 // constant (which, since the types must match, means that it is a 1446 // ConstantPointerNull). 1447 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) { 1448 // Block address in another function can't equal this one, but block 1449 // addresses in the current function might be the same if blocks are 1450 // empty. 1451 if (BA2->getFunction() != BA->getFunction()) 1452 return ICmpInst::ICMP_NE; 1453 } else { 1454 // Block addresses aren't null, don't equal the address of globals. 1455 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) && 1456 "Canonicalization guarantee!"); 1457 return ICmpInst::ICMP_NE; 1458 } 1459 } else { 1460 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a 1461 // constantexpr, a global, block address, or a simple constant. 1462 ConstantExpr *CE1 = cast<ConstantExpr>(V1); 1463 Constant *CE1Op0 = CE1->getOperand(0); 1464 1465 switch (CE1->getOpcode()) { 1466 case Instruction::Trunc: 1467 case Instruction::FPTrunc: 1468 case Instruction::FPExt: 1469 case Instruction::FPToUI: 1470 case Instruction::FPToSI: 1471 break; // We can't evaluate floating point casts or truncations. 1472 1473 case Instruction::BitCast: 1474 // If this is a global value cast, check to see if the RHS is also a 1475 // GlobalValue. 1476 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) 1477 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) 1478 return areGlobalsPotentiallyEqual(GV, GV2); 1479 [[fallthrough]]; 1480 case Instruction::UIToFP: 1481 case Instruction::SIToFP: 1482 case Instruction::ZExt: 1483 case Instruction::SExt: 1484 // We can't evaluate floating point casts or truncations. 1485 if (CE1Op0->getType()->isFPOrFPVectorTy()) 1486 break; 1487 1488 // If the cast is not actually changing bits, and the second operand is a 1489 // null pointer, do the comparison with the pre-casted value. 1490 if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) { 1491 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false; 1492 if (CE1->getOpcode() == Instruction::SExt) isSigned = true; 1493 return evaluateICmpRelation(CE1Op0, 1494 Constant::getNullValue(CE1Op0->getType()), 1495 isSigned); 1496 } 1497 break; 1498 1499 case Instruction::GetElementPtr: { 1500 GEPOperator *CE1GEP = cast<GEPOperator>(CE1); 1501 // Ok, since this is a getelementptr, we know that the constant has a 1502 // pointer type. Check the various cases. 1503 if (isa<ConstantPointerNull>(V2)) { 1504 // If we are comparing a GEP to a null pointer, check to see if the base 1505 // of the GEP equals the null pointer. 1506 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1507 // If its not weak linkage, the GVal must have a non-zero address 1508 // so the result is greater-than 1509 if (!GV->hasExternalWeakLinkage() && CE1GEP->isInBounds()) 1510 return ICmpInst::ICMP_UGT; 1511 } 1512 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) { 1513 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) { 1514 if (GV != GV2) { 1515 if (CE1GEP->hasAllZeroIndices()) 1516 return areGlobalsPotentiallyEqual(GV, GV2); 1517 return ICmpInst::BAD_ICMP_PREDICATE; 1518 } 1519 } 1520 } else if (const auto *CE2GEP = dyn_cast<GEPOperator>(V2)) { 1521 // By far the most common case to handle is when the base pointers are 1522 // obviously to the same global. 1523 const Constant *CE2Op0 = cast<Constant>(CE2GEP->getPointerOperand()); 1524 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) { 1525 // Don't know relative ordering, but check for inequality. 1526 if (CE1Op0 != CE2Op0) { 1527 if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) 1528 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0), 1529 cast<GlobalValue>(CE2Op0)); 1530 return ICmpInst::BAD_ICMP_PREDICATE; 1531 } 1532 } 1533 } 1534 break; 1535 } 1536 default: 1537 break; 1538 } 1539 } 1540 1541 return ICmpInst::BAD_ICMP_PREDICATE; 1542 } 1543 1544 static Constant *constantFoldCompareGlobalToNull(CmpInst::Predicate Predicate, 1545 Constant *C1, Constant *C2) { 1546 const GlobalValue *GV = dyn_cast<GlobalValue>(C2); 1547 if (!GV || !C1->isNullValue()) 1548 return nullptr; 1549 1550 // Don't try to evaluate aliases. External weak GV can be null. 1551 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() && 1552 !NullPointerIsDefined(nullptr /* F */, 1553 GV->getType()->getAddressSpace())) { 1554 if (Predicate == ICmpInst::ICMP_EQ) 1555 return ConstantInt::getFalse(C1->getContext()); 1556 else if (Predicate == ICmpInst::ICMP_NE) 1557 return ConstantInt::getTrue(C1->getContext()); 1558 } 1559 1560 return nullptr; 1561 } 1562 1563 Constant *llvm::ConstantFoldCompareInstruction(CmpInst::Predicate Predicate, 1564 Constant *C1, Constant *C2) { 1565 Type *ResultTy; 1566 if (VectorType *VT = dyn_cast<VectorType>(C1->getType())) 1567 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()), 1568 VT->getElementCount()); 1569 else 1570 ResultTy = Type::getInt1Ty(C1->getContext()); 1571 1572 // Fold FCMP_FALSE/FCMP_TRUE unconditionally. 1573 if (Predicate == FCmpInst::FCMP_FALSE) 1574 return Constant::getNullValue(ResultTy); 1575 1576 if (Predicate == FCmpInst::FCMP_TRUE) 1577 return Constant::getAllOnesValue(ResultTy); 1578 1579 // Handle some degenerate cases first 1580 if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2)) 1581 return PoisonValue::get(ResultTy); 1582 1583 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) { 1584 bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate); 1585 // For EQ and NE, we can always pick a value for the undef to make the 1586 // predicate pass or fail, so we can return undef. 1587 // Also, if both operands are undef, we can return undef for int comparison. 1588 if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2)) 1589 return UndefValue::get(ResultTy); 1590 1591 // Otherwise, for integer compare, pick the same value as the non-undef 1592 // operand, and fold it to true or false. 1593 if (isIntegerPredicate) 1594 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate)); 1595 1596 // Choosing NaN for the undef will always make unordered comparison succeed 1597 // and ordered comparison fails. 1598 return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate)); 1599 } 1600 1601 // icmp eq/ne(null,GV) -> false/true 1602 if (Constant *Folded = constantFoldCompareGlobalToNull(Predicate, C1, C2)) 1603 return Folded; 1604 1605 // icmp eq/ne(GV,null) -> false/true 1606 if (Constant *Folded = constantFoldCompareGlobalToNull(Predicate, C2, C1)) 1607 return Folded; 1608 1609 if (C2->isNullValue()) { 1610 // The caller is expected to commute the operands if the constant expression 1611 // is C2. 1612 // C1 >= 0 --> true 1613 if (Predicate == ICmpInst::ICMP_UGE) 1614 return Constant::getAllOnesValue(ResultTy); 1615 // C1 < 0 --> false 1616 if (Predicate == ICmpInst::ICMP_ULT) 1617 return Constant::getNullValue(ResultTy); 1618 } 1619 1620 // If the comparison is a comparison between two i1's, simplify it. 1621 if (C1->getType()->isIntegerTy(1)) { 1622 switch (Predicate) { 1623 case ICmpInst::ICMP_EQ: 1624 if (isa<ConstantInt>(C2)) 1625 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2)); 1626 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2); 1627 case ICmpInst::ICMP_NE: 1628 return ConstantExpr::getXor(C1, C2); 1629 default: 1630 break; 1631 } 1632 } 1633 1634 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) { 1635 const APInt &V1 = cast<ConstantInt>(C1)->getValue(); 1636 const APInt &V2 = cast<ConstantInt>(C2)->getValue(); 1637 return ConstantInt::get(ResultTy, ICmpInst::compare(V1, V2, Predicate)); 1638 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) { 1639 const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF(); 1640 const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF(); 1641 return ConstantInt::get(ResultTy, FCmpInst::compare(C1V, C2V, Predicate)); 1642 } else if (auto *C1VTy = dyn_cast<VectorType>(C1->getType())) { 1643 1644 // Fast path for splatted constants. 1645 if (Constant *C1Splat = C1->getSplatValue()) 1646 if (Constant *C2Splat = C2->getSplatValue()) 1647 return ConstantVector::getSplat( 1648 C1VTy->getElementCount(), 1649 ConstantExpr::getCompare(Predicate, C1Splat, C2Splat)); 1650 1651 // Do not iterate on scalable vector. The number of elements is unknown at 1652 // compile-time. 1653 if (isa<ScalableVectorType>(C1VTy)) 1654 return nullptr; 1655 1656 // If we can constant fold the comparison of each element, constant fold 1657 // the whole vector comparison. 1658 SmallVector<Constant*, 4> ResElts; 1659 Type *Ty = IntegerType::get(C1->getContext(), 32); 1660 // Compare the elements, producing an i1 result or constant expr. 1661 for (unsigned I = 0, E = C1VTy->getElementCount().getKnownMinValue(); 1662 I != E; ++I) { 1663 Constant *C1E = 1664 ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, I)); 1665 Constant *C2E = 1666 ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, I)); 1667 1668 ResElts.push_back(ConstantExpr::getCompare(Predicate, C1E, C2E)); 1669 } 1670 1671 return ConstantVector::get(ResElts); 1672 } 1673 1674 if (C1->getType()->isFloatingPointTy() && 1675 // Only call evaluateFCmpRelation if we have a constant expr to avoid 1676 // infinite recursive loop 1677 (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) { 1678 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1679 switch (evaluateFCmpRelation(C1, C2)) { 1680 default: llvm_unreachable("Unknown relation!"); 1681 case FCmpInst::FCMP_UNO: 1682 case FCmpInst::FCMP_ORD: 1683 case FCmpInst::FCMP_UNE: 1684 case FCmpInst::FCMP_ULT: 1685 case FCmpInst::FCMP_UGT: 1686 case FCmpInst::FCMP_ULE: 1687 case FCmpInst::FCMP_UGE: 1688 case FCmpInst::FCMP_TRUE: 1689 case FCmpInst::FCMP_FALSE: 1690 case FCmpInst::BAD_FCMP_PREDICATE: 1691 break; // Couldn't determine anything about these constants. 1692 case FCmpInst::FCMP_OEQ: // We know that C1 == C2 1693 Result = 1694 (Predicate == FCmpInst::FCMP_UEQ || Predicate == FCmpInst::FCMP_OEQ || 1695 Predicate == FCmpInst::FCMP_ULE || Predicate == FCmpInst::FCMP_OLE || 1696 Predicate == FCmpInst::FCMP_UGE || Predicate == FCmpInst::FCMP_OGE); 1697 break; 1698 case FCmpInst::FCMP_OLT: // We know that C1 < C2 1699 Result = 1700 (Predicate == FCmpInst::FCMP_UNE || Predicate == FCmpInst::FCMP_ONE || 1701 Predicate == FCmpInst::FCMP_ULT || Predicate == FCmpInst::FCMP_OLT || 1702 Predicate == FCmpInst::FCMP_ULE || Predicate == FCmpInst::FCMP_OLE); 1703 break; 1704 case FCmpInst::FCMP_OGT: // We know that C1 > C2 1705 Result = 1706 (Predicate == FCmpInst::FCMP_UNE || Predicate == FCmpInst::FCMP_ONE || 1707 Predicate == FCmpInst::FCMP_UGT || Predicate == FCmpInst::FCMP_OGT || 1708 Predicate == FCmpInst::FCMP_UGE || Predicate == FCmpInst::FCMP_OGE); 1709 break; 1710 case FCmpInst::FCMP_OLE: // We know that C1 <= C2 1711 // We can only partially decide this relation. 1712 if (Predicate == FCmpInst::FCMP_UGT || Predicate == FCmpInst::FCMP_OGT) 1713 Result = 0; 1714 else if (Predicate == FCmpInst::FCMP_ULT || 1715 Predicate == FCmpInst::FCMP_OLT) 1716 Result = 1; 1717 break; 1718 case FCmpInst::FCMP_OGE: // We known that C1 >= C2 1719 // We can only partially decide this relation. 1720 if (Predicate == FCmpInst::FCMP_ULT || Predicate == FCmpInst::FCMP_OLT) 1721 Result = 0; 1722 else if (Predicate == FCmpInst::FCMP_UGT || 1723 Predicate == FCmpInst::FCMP_OGT) 1724 Result = 1; 1725 break; 1726 case FCmpInst::FCMP_ONE: // We know that C1 != C2 1727 // We can only partially decide this relation. 1728 if (Predicate == FCmpInst::FCMP_OEQ || Predicate == FCmpInst::FCMP_UEQ) 1729 Result = 0; 1730 else if (Predicate == FCmpInst::FCMP_ONE || 1731 Predicate == FCmpInst::FCMP_UNE) 1732 Result = 1; 1733 break; 1734 case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2). 1735 // We can only partially decide this relation. 1736 if (Predicate == FCmpInst::FCMP_ONE) 1737 Result = 0; 1738 else if (Predicate == FCmpInst::FCMP_UEQ) 1739 Result = 1; 1740 break; 1741 } 1742 1743 // If we evaluated the result, return it now. 1744 if (Result != -1) 1745 return ConstantInt::get(ResultTy, Result); 1746 1747 } else { 1748 // Evaluate the relation between the two constants, per the predicate. 1749 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. 1750 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(Predicate))) { 1751 default: llvm_unreachable("Unknown relational!"); 1752 case ICmpInst::BAD_ICMP_PREDICATE: 1753 break; // Couldn't determine anything about these constants. 1754 case ICmpInst::ICMP_EQ: // We know the constants are equal! 1755 // If we know the constants are equal, we can decide the result of this 1756 // computation precisely. 1757 Result = ICmpInst::isTrueWhenEqual(Predicate); 1758 break; 1759 case ICmpInst::ICMP_ULT: 1760 switch (Predicate) { 1761 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: 1762 Result = 1; break; 1763 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: 1764 Result = 0; break; 1765 default: 1766 break; 1767 } 1768 break; 1769 case ICmpInst::ICMP_SLT: 1770 switch (Predicate) { 1771 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: 1772 Result = 1; break; 1773 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: 1774 Result = 0; break; 1775 default: 1776 break; 1777 } 1778 break; 1779 case ICmpInst::ICMP_UGT: 1780 switch (Predicate) { 1781 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: 1782 Result = 1; break; 1783 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: 1784 Result = 0; break; 1785 default: 1786 break; 1787 } 1788 break; 1789 case ICmpInst::ICMP_SGT: 1790 switch (Predicate) { 1791 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: 1792 Result = 1; break; 1793 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: 1794 Result = 0; break; 1795 default: 1796 break; 1797 } 1798 break; 1799 case ICmpInst::ICMP_ULE: 1800 if (Predicate == ICmpInst::ICMP_UGT) 1801 Result = 0; 1802 if (Predicate == ICmpInst::ICMP_ULT || Predicate == ICmpInst::ICMP_ULE) 1803 Result = 1; 1804 break; 1805 case ICmpInst::ICMP_SLE: 1806 if (Predicate == ICmpInst::ICMP_SGT) 1807 Result = 0; 1808 if (Predicate == ICmpInst::ICMP_SLT || Predicate == ICmpInst::ICMP_SLE) 1809 Result = 1; 1810 break; 1811 case ICmpInst::ICMP_UGE: 1812 if (Predicate == ICmpInst::ICMP_ULT) 1813 Result = 0; 1814 if (Predicate == ICmpInst::ICMP_UGT || Predicate == ICmpInst::ICMP_UGE) 1815 Result = 1; 1816 break; 1817 case ICmpInst::ICMP_SGE: 1818 if (Predicate == ICmpInst::ICMP_SLT) 1819 Result = 0; 1820 if (Predicate == ICmpInst::ICMP_SGT || Predicate == ICmpInst::ICMP_SGE) 1821 Result = 1; 1822 break; 1823 case ICmpInst::ICMP_NE: 1824 if (Predicate == ICmpInst::ICMP_EQ) 1825 Result = 0; 1826 if (Predicate == ICmpInst::ICMP_NE) 1827 Result = 1; 1828 break; 1829 } 1830 1831 // If we evaluated the result, return it now. 1832 if (Result != -1) 1833 return ConstantInt::get(ResultTy, Result); 1834 1835 // If the right hand side is a bitcast, try using its inverse to simplify 1836 // it by moving it to the left hand side. We can't do this if it would turn 1837 // a vector compare into a scalar compare or visa versa, or if it would turn 1838 // the operands into FP values. 1839 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) { 1840 Constant *CE2Op0 = CE2->getOperand(0); 1841 if (CE2->getOpcode() == Instruction::BitCast && 1842 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() && 1843 !CE2Op0->getType()->isFPOrFPVectorTy()) { 1844 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType()); 1845 return ConstantExpr::getICmp(Predicate, Inverse, CE2Op0); 1846 } 1847 } 1848 1849 // If the left hand side is an extension, try eliminating it. 1850 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) { 1851 if ((CE1->getOpcode() == Instruction::SExt && 1852 ICmpInst::isSigned(Predicate)) || 1853 (CE1->getOpcode() == Instruction::ZExt && 1854 !ICmpInst::isSigned(Predicate))) { 1855 Constant *CE1Op0 = CE1->getOperand(0); 1856 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType()); 1857 if (CE1Inverse == CE1Op0) { 1858 // Check whether we can safely truncate the right hand side. 1859 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType()); 1860 if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse, 1861 C2->getType()) == C2) 1862 return ConstantExpr::getICmp(Predicate, CE1Inverse, C2Inverse); 1863 } 1864 } 1865 } 1866 1867 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) || 1868 (C1->isNullValue() && !C2->isNullValue())) { 1869 // If C2 is a constant expr and C1 isn't, flip them around and fold the 1870 // other way if possible. 1871 // Also, if C1 is null and C2 isn't, flip them around. 1872 Predicate = ICmpInst::getSwappedPredicate(Predicate); 1873 return ConstantExpr::getICmp(Predicate, C2, C1); 1874 } 1875 } 1876 return nullptr; 1877 } 1878 1879 /// Test whether the given sequence of *normalized* indices is "inbounds". 1880 template<typename IndexTy> 1881 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) { 1882 // No indices means nothing that could be out of bounds. 1883 if (Idxs.empty()) return true; 1884 1885 // If the first index is zero, it's in bounds. 1886 if (cast<Constant>(Idxs[0])->isNullValue()) return true; 1887 1888 // If the first index is one and all the rest are zero, it's in bounds, 1889 // by the one-past-the-end rule. 1890 if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) { 1891 if (!CI->isOne()) 1892 return false; 1893 } else { 1894 auto *CV = cast<ConstantDataVector>(Idxs[0]); 1895 CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue()); 1896 if (!CI || !CI->isOne()) 1897 return false; 1898 } 1899 1900 for (unsigned i = 1, e = Idxs.size(); i != e; ++i) 1901 if (!cast<Constant>(Idxs[i])->isNullValue()) 1902 return false; 1903 return true; 1904 } 1905 1906 /// Test whether a given ConstantInt is in-range for a SequentialType. 1907 static bool isIndexInRangeOfArrayType(uint64_t NumElements, 1908 const ConstantInt *CI) { 1909 // We cannot bounds check the index if it doesn't fit in an int64_t. 1910 if (CI->getValue().getSignificantBits() > 64) 1911 return false; 1912 1913 // A negative index or an index past the end of our sequential type is 1914 // considered out-of-range. 1915 int64_t IndexVal = CI->getSExtValue(); 1916 if (IndexVal < 0 || (IndexVal != 0 && (uint64_t)IndexVal >= NumElements)) 1917 return false; 1918 1919 // Otherwise, it is in-range. 1920 return true; 1921 } 1922 1923 // Combine Indices - If the source pointer to this getelementptr instruction 1924 // is a getelementptr instruction, combine the indices of the two 1925 // getelementptr instructions into a single instruction. 1926 static Constant *foldGEPOfGEP(GEPOperator *GEP, Type *PointeeTy, bool InBounds, 1927 ArrayRef<Value *> Idxs) { 1928 if (PointeeTy != GEP->getResultElementType()) 1929 return nullptr; 1930 1931 Constant *Idx0 = cast<Constant>(Idxs[0]); 1932 if (Idx0->isNullValue()) { 1933 // Handle the simple case of a zero index. 1934 SmallVector<Value*, 16> NewIndices; 1935 NewIndices.reserve(Idxs.size() + GEP->getNumIndices()); 1936 NewIndices.append(GEP->idx_begin(), GEP->idx_end()); 1937 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 1938 return ConstantExpr::getGetElementPtr( 1939 GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()), 1940 NewIndices, InBounds && GEP->isInBounds(), GEP->getInRangeIndex()); 1941 } 1942 1943 gep_type_iterator LastI = gep_type_end(GEP); 1944 for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP); 1945 I != E; ++I) 1946 LastI = I; 1947 1948 // We can't combine GEPs if the last index is a struct type. 1949 if (!LastI.isSequential()) 1950 return nullptr; 1951 // We could perform the transform with non-constant index, but prefer leaving 1952 // it as GEP of GEP rather than GEP of add for now. 1953 ConstantInt *CI = dyn_cast<ConstantInt>(Idx0); 1954 if (!CI) 1955 return nullptr; 1956 1957 // TODO: This code may be extended to handle vectors as well. 1958 auto *LastIdx = cast<Constant>(GEP->getOperand(GEP->getNumOperands()-1)); 1959 Type *LastIdxTy = LastIdx->getType(); 1960 if (LastIdxTy->isVectorTy()) 1961 return nullptr; 1962 1963 SmallVector<Value*, 16> NewIndices; 1964 NewIndices.reserve(Idxs.size() + GEP->getNumIndices()); 1965 NewIndices.append(GEP->idx_begin(), GEP->idx_end() - 1); 1966 1967 // Add the last index of the source with the first index of the new GEP. 1968 // Make sure to handle the case when they are actually different types. 1969 if (LastIdxTy != Idx0->getType()) { 1970 unsigned CommonExtendedWidth = 1971 std::max(LastIdxTy->getIntegerBitWidth(), 1972 Idx0->getType()->getIntegerBitWidth()); 1973 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 1974 1975 Type *CommonTy = 1976 Type::getIntNTy(LastIdxTy->getContext(), CommonExtendedWidth); 1977 Idx0 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy); 1978 LastIdx = ConstantExpr::getSExtOrBitCast(LastIdx, CommonTy); 1979 } 1980 1981 NewIndices.push_back(ConstantExpr::get(Instruction::Add, Idx0, LastIdx)); 1982 NewIndices.append(Idxs.begin() + 1, Idxs.end()); 1983 1984 // The combined GEP normally inherits its index inrange attribute from 1985 // the inner GEP, but if the inner GEP's last index was adjusted by the 1986 // outer GEP, any inbounds attribute on that index is invalidated. 1987 std::optional<unsigned> IRIndex = GEP->getInRangeIndex(); 1988 if (IRIndex && *IRIndex == GEP->getNumIndices() - 1) 1989 IRIndex = std::nullopt; 1990 1991 return ConstantExpr::getGetElementPtr( 1992 GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()), 1993 NewIndices, InBounds && GEP->isInBounds(), IRIndex); 1994 } 1995 1996 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, 1997 bool InBounds, 1998 std::optional<unsigned> InRangeIndex, 1999 ArrayRef<Value *> Idxs) { 2000 if (Idxs.empty()) return C; 2001 2002 Type *GEPTy = GetElementPtrInst::getGEPReturnType( 2003 C, ArrayRef((Value *const *)Idxs.data(), Idxs.size())); 2004 2005 if (isa<PoisonValue>(C)) 2006 return PoisonValue::get(GEPTy); 2007 2008 if (isa<UndefValue>(C)) 2009 // If inbounds, we can choose an out-of-bounds pointer as a base pointer. 2010 return InBounds ? PoisonValue::get(GEPTy) : UndefValue::get(GEPTy); 2011 2012 auto IsNoOp = [&]() { 2013 // Avoid losing inrange information. 2014 if (InRangeIndex) 2015 return false; 2016 2017 return all_of(Idxs, [](Value *Idx) { 2018 Constant *IdxC = cast<Constant>(Idx); 2019 return IdxC->isNullValue() || isa<UndefValue>(IdxC); 2020 }); 2021 }; 2022 if (IsNoOp()) 2023 return GEPTy->isVectorTy() && !C->getType()->isVectorTy() 2024 ? ConstantVector::getSplat( 2025 cast<VectorType>(GEPTy)->getElementCount(), C) 2026 : C; 2027 2028 if (C->isNullValue()) { 2029 bool isNull = true; 2030 for (Value *Idx : Idxs) 2031 if (!isa<UndefValue>(Idx) && !cast<Constant>(Idx)->isNullValue()) { 2032 isNull = false; 2033 break; 2034 } 2035 if (isNull) { 2036 PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType()); 2037 Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs); 2038 2039 assert(Ty && "Invalid indices for GEP!"); 2040 Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2041 Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace()); 2042 if (VectorType *VT = dyn_cast<VectorType>(C->getType())) 2043 GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount()); 2044 2045 // The GEP returns a vector of pointers when one of more of 2046 // its arguments is a vector. 2047 for (Value *Idx : Idxs) { 2048 if (auto *VT = dyn_cast<VectorType>(Idx->getType())) { 2049 assert((!isa<VectorType>(GEPTy) || isa<ScalableVectorType>(GEPTy) == 2050 isa<ScalableVectorType>(VT)) && 2051 "Mismatched GEPTy vector types"); 2052 GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount()); 2053 break; 2054 } 2055 } 2056 2057 return Constant::getNullValue(GEPTy); 2058 } 2059 } 2060 2061 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2062 if (auto *GEP = dyn_cast<GEPOperator>(CE)) 2063 if (Constant *C = foldGEPOfGEP(GEP, PointeeTy, InBounds, Idxs)) 2064 return C; 2065 2066 // Check to see if any array indices are not within the corresponding 2067 // notional array or vector bounds. If so, try to determine if they can be 2068 // factored out into preceding dimensions. 2069 SmallVector<Constant *, 8> NewIdxs; 2070 Type *Ty = PointeeTy; 2071 Type *Prev = C->getType(); 2072 auto GEPIter = gep_type_begin(PointeeTy, Idxs); 2073 bool Unknown = 2074 !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]); 2075 for (unsigned i = 1, e = Idxs.size(); i != e; 2076 Prev = Ty, Ty = (++GEPIter).getIndexedType(), ++i) { 2077 if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) { 2078 // We don't know if it's in range or not. 2079 Unknown = true; 2080 continue; 2081 } 2082 if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1])) 2083 // Skip if the type of the previous index is not supported. 2084 continue; 2085 if (InRangeIndex && i == *InRangeIndex + 1) { 2086 // If an index is marked inrange, we cannot apply this canonicalization to 2087 // the following index, as that will cause the inrange index to point to 2088 // the wrong element. 2089 continue; 2090 } 2091 if (isa<StructType>(Ty)) { 2092 // The verify makes sure that GEPs into a struct are in range. 2093 continue; 2094 } 2095 if (isa<VectorType>(Ty)) { 2096 // There can be awkward padding in after a non-power of two vector. 2097 Unknown = true; 2098 continue; 2099 } 2100 auto *STy = cast<ArrayType>(Ty); 2101 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) { 2102 if (isIndexInRangeOfArrayType(STy->getNumElements(), CI)) 2103 // It's in range, skip to the next index. 2104 continue; 2105 if (CI->isNegative()) { 2106 // It's out of range and negative, don't try to factor it. 2107 Unknown = true; 2108 continue; 2109 } 2110 } else { 2111 auto *CV = cast<ConstantDataVector>(Idxs[i]); 2112 bool InRange = true; 2113 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) { 2114 auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I)); 2115 InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI); 2116 if (CI->isNegative()) { 2117 Unknown = true; 2118 break; 2119 } 2120 } 2121 if (InRange || Unknown) 2122 // It's in range, skip to the next index. 2123 // It's out of range and negative, don't try to factor it. 2124 continue; 2125 } 2126 if (isa<StructType>(Prev)) { 2127 // It's out of range, but the prior dimension is a struct 2128 // so we can't do anything about it. 2129 Unknown = true; 2130 continue; 2131 } 2132 2133 // Determine the number of elements in our sequential type. 2134 uint64_t NumElements = STy->getArrayNumElements(); 2135 if (!NumElements) { 2136 Unknown = true; 2137 continue; 2138 } 2139 2140 // It's out of range, but we can factor it into the prior 2141 // dimension. 2142 NewIdxs.resize(Idxs.size()); 2143 2144 // Expand the current index or the previous index to a vector from a scalar 2145 // if necessary. 2146 Constant *CurrIdx = cast<Constant>(Idxs[i]); 2147 auto *PrevIdx = 2148 NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]); 2149 bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy(); 2150 bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy(); 2151 bool UseVector = IsCurrIdxVector || IsPrevIdxVector; 2152 2153 if (!IsCurrIdxVector && IsPrevIdxVector) 2154 CurrIdx = ConstantDataVector::getSplat( 2155 cast<FixedVectorType>(PrevIdx->getType())->getNumElements(), CurrIdx); 2156 2157 if (!IsPrevIdxVector && IsCurrIdxVector) 2158 PrevIdx = ConstantDataVector::getSplat( 2159 cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), PrevIdx); 2160 2161 Constant *Factor = 2162 ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements); 2163 if (UseVector) 2164 Factor = ConstantDataVector::getSplat( 2165 IsPrevIdxVector 2166 ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements() 2167 : cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), 2168 Factor); 2169 2170 NewIdxs[i] = 2171 ConstantFoldBinaryInstruction(Instruction::SRem, CurrIdx, Factor); 2172 2173 Constant *Div = 2174 ConstantFoldBinaryInstruction(Instruction::SDiv, CurrIdx, Factor); 2175 2176 // We're working on either ConstantInt or vectors of ConstantInt, 2177 // so these should always fold. 2178 assert(NewIdxs[i] != nullptr && Div != nullptr && "Should have folded"); 2179 2180 unsigned CommonExtendedWidth = 2181 std::max(PrevIdx->getType()->getScalarSizeInBits(), 2182 Div->getType()->getScalarSizeInBits()); 2183 CommonExtendedWidth = std::max(CommonExtendedWidth, 64U); 2184 2185 // Before adding, extend both operands to i64 to avoid 2186 // overflow trouble. 2187 Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth); 2188 if (UseVector) 2189 ExtendedTy = FixedVectorType::get( 2190 ExtendedTy, 2191 IsPrevIdxVector 2192 ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements() 2193 : cast<FixedVectorType>(CurrIdx->getType())->getNumElements()); 2194 2195 if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2196 PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy); 2197 2198 if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth)) 2199 Div = ConstantExpr::getSExt(Div, ExtendedTy); 2200 2201 NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div); 2202 } 2203 2204 // If we did any factoring, start over with the adjusted indices. 2205 if (!NewIdxs.empty()) { 2206 for (unsigned i = 0, e = Idxs.size(); i != e; ++i) 2207 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]); 2208 return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds, 2209 InRangeIndex); 2210 } 2211 2212 // If all indices are known integers and normalized, we can do a simple 2213 // check for the "inbounds" property. 2214 if (!Unknown && !InBounds) 2215 if (auto *GV = dyn_cast<GlobalVariable>(C)) 2216 if (!GV->hasExternalWeakLinkage() && GV->getValueType() == PointeeTy && 2217 isInBoundsIndices(Idxs)) 2218 return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs, 2219 /*InBounds=*/true, InRangeIndex); 2220 2221 return nullptr; 2222 } 2223